📈 economics

Advances in Agentic AI: Back to the Future

This paper establishes a structured analytical framework for Agentic AI by defining key concepts, tracing the evolution of Algorithmization methodologies, and introducing a two-tiered "Machine in Machine Learning" (M1 and M2) model to distinguish current LLM-based retrieval systems from the future of production-grade, strategies-based B2B transformation.

Original authors: Sergio Alvarez-Telena, Marta Diez-Fernandez

Published 2026-03-26

📖 7 min read🧠 Deep dive

CC BY 4.0

Original authors: Sergio Alvarez-Telena, Marta Diez-Fernandez

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The "Brain" vs. The "Body"

Imagine you are building a super-intelligent robot.

The "Learning" (L): This is the robot's brain. It's the part that reads books, learns math, and tries to understand the world. The paper argues that this part is becoming a "commodity." Just like everyone can buy a powerful CPU chip today, everyone can access powerful AI brains (like Large Language Models). Having the smartest brain doesn't guarantee you win the race anymore.
The "Machine" (M): This is the robot's body, nervous system, and hands. It's how the brain actually does things in the real world. It's how it follows rules, talks to other robots, keeps secrets, and doesn't crash when things go wrong.

The Paper's Main Point:
Most companies are currently obsessed with buying the "smartest brain" (LLMs) and hoping it will fix their business. The authors say this is a mistake. The real magic isn't in the brain; it's in building a body that can actually use that brain to run a complex company. They call this Strategies-based Agentic AI.

The Two Types of "Machines" (M1 vs. M2)

The authors split the "Machine" part into two levels:

1. M1: The "Model Factory" (The Brain Builder)

What it is: This is the heavy machinery needed to train the AI brain. It involves massive computers, huge amounts of data, and complex math to create the model.
The Analogy: Think of M1 as a high-tech bakery that bakes the perfect loaf of bread.
The Problem: Many companies think if they buy the bread (the AI model), they can run a restaurant. But baking bread is different from running a restaurant. The bakery (M1) is great at making the product, but it doesn't know how to serve customers, manage the menu, or handle the health inspector.
Current Trend: Most "Agentic AI" today is just M1. It's a fancy chatbot that can write code or answer questions, but it's not built to run a whole corporation safely.

2. M2: The "Operating System" (The Body Builder)

What it is: This is the complex architecture that takes the AI brain and plugs it into a real business. It connects the AI to HR, finance, cybersecurity, and legal teams. It ensures the AI follows the rules and doesn't make up facts (hallucinate).
The Analogy: Think of M2 as the entire restaurant ecosystem. It's the kitchen layout, the waiters, the reservation system, the health code compliance, and the manager who knows when to fire a waiter or change the menu.
The Goal: M2 is what turns a "cool toy" into a "production-grade business engine." It allows a company to be "AI-first," meaning the whole organization runs like a coordinated machine, not just a person typing into a chat box.

Why Are So Many AI Projects Failing? (The 95% Failure Rate)

The paper explains that 95% of AI projects fail not because the AI is "dumb," but because companies are trying to put a square peg in a round hole.

The "Vibe Coding" Trap: Many companies are trying to build their business on "vibe coding" (just asking an AI to "make it work" without a solid plan).
- Analogy: It's like trying to build a skyscraper by asking a genie to "make it tall." The genie might make a tall pile of bricks, but it won't have plumbing, electricity, or a foundation. It will collapse.
The "Hallucination" Misunderstanding: People think AI "hallucinating" (making things up) is a bug that will be fixed. The authors say it's a feature of the design. Because LLMs guess the next word based on probability, they will make things up. If you build a bank or a hospital on a system that guesses, you are building on sand.
The Talent Mix-up: Companies are hiring "Auditors" (people who check boxes) to lead "Transformations" (people who build new things).
- Analogy: It's like hiring a health inspector to design a new Ferrari. The inspector is great at checking if the car is safe, but they don't know how to build an engine. You need a different kind of talent to drive the transformation.

The Solution: "Strategies-Based" AI

Instead of just letting the AI talk, the authors propose a system where the AI is guided by human strategy and strict rules.

The "Augmented Machine": Imagine a pilot and a co-pilot. The human pilot (the expert) sets the destination and the safety rules. The AI co-pilot (the machine) flies the plane, handles the turbulence, and manages the fuel.
The "Federated" Approach: Instead of one giant, scary AI controlling everything, the system is made of many small, specialized "Smart Agents."
- Analogy: Think of a hive of bees. No single bee knows the whole plan. One bee checks the flowers, another checks the weather, another guards the hive. They talk to each other and work together. If one bee gets confused, the hive doesn't crash. This is safer and more flexible than one giant "super-bee."

The "Back to the Future" Twist

The title "Back to the Future" is a joke about how we are going in circles.

The Past: In the 1990s and 2000s, experts in finance (like the authors) built complex, rule-based systems to trade stocks. They knew that "smart models" weren't enough; you needed a robust structure.
The Present: We are back to the same problem, but now we have "AI" instead of "spreadsheets." We are making the same mistake of thinking the "smart model" is the solution, forgetting the "structure."
The Future: The paper argues we need to go back to the old-school discipline of building solid, rule-based structures (M2) and then plug the new AI brains into them.

The "Secret Sauce" (Why This is Hard)

The authors admit that building this M2 system is incredibly difficult. It's not just coding; it's a mix of:

Psychology: Getting people to change how they work.
Law: Making sure the AI doesn't break regulations.
Security: Making sure hackers can't trick the AI.
Economics: Making sure the whole thing actually makes money.

They built this system over 10 years in a "Centre of Excellence" (a secret lab), testing it on high-stakes things like algorithmic trading (where one mistake costs millions) and cybersecurity. Because they tested it in the "danger zone" first, they know it works.

Summary for the Everyday Person

Don't just buy a smart AI brain.
If you want to use AI to run your business, country, or life, you can't just ask a chatbot to "do it." You need to build a digital body (M2) that knows how to walk, talk, follow laws, and protect itself.

M1 (The Brain): Everyone has access to this. It's the "commodity."
M2 (The Body): This is the real competitive advantage. It's the custom-built, rule-following, secure system that actually gets the job done.

The future belongs to the companies that stop trying to "prompt engineer" their way to success and start architecting their way to a new kind of organization where humans and machines work together like a well-oiled machine.

1. Problem Statement

The paper identifies a critical conceptual and architectural misalignment in the current discourse surrounding Agentic AI. The authors argue that the industry is suffering from a "perfect storm" of misconceptions driven by:

Conflation of Terms: The confusion between "Science" (model creation), "Science Applied" (model calibration), and "Applied Science" (operational deployment).
LLM-Centric Fallacy: The widespread belief that Agentic AI is synonymous with Large Language Models (LLMs). The authors contend that LLMs are statistical estimators prone to structural "hallucinations" (errors by design) and are unsuitable as the backbone for production-grade, mission-critical B2B transformation.
Organizational Inertia: Most corporations are attempting to layer AI onto legacy operational structures rather than transforming their underlying architecture. This leads to the cited statistic that 95% of AI projects fail due to structural incompatibility, not technological inadequacy.
Talent and Role Mismanagement: The confusion between "Catalyzers" (architects who drive transformation) and "Auditors" (who merely observe), leading to strategic stagnation.

2. Methodology and Theoretical Framework

The paper proposes a new theoretical framework called Algorithmization, which redefines Machine Learning (ML) by distinguishing between its two core components: Learning (L) and Machine (M).

A. The Machine Theory of Agentic AI (M1 vs. M2)

The authors decompose ML into two distinct "Machines":

M1 (The Calibration Machine): Focuses on the Learning (L) component. It involves the data engineering, hardware management, and iterative optimization required to estimate the statistical model $f(x)$ $f (x)$ .
- Current State: Dominated by LLMs and academic research. The authors argue $f(x)$ is becoming a commodity; competitive advantage cannot be sustained here.
- Limitation: M1 systems (like LLMs) are often black-box, stochastic, and lack the deterministic rigor required for enterprise operations.
M2 (The Consumption Machine): Focuses on the Applied Science component. It is the architectural discipline required to consume the model $f(x)$ $f (x)$ and integrate it into a federated, strategic, and operational ecosystem.
- Definition: M2 is a Strategies-based Agentic AI system. It orchestrates "if-then" logic, expert heuristics, compliance constraints, and data flows across an organization.
- Key Distinction: Unlike LLM-based agents (which rely on "vibe coding" and probabilistic generation), M2 agents are deterministic, modular, and designed for production-grade reliability.

B. Architectural Constructs

The paper introduces specific architectural units for M2:

Minimal Architecture Units (MAUs): Core computational entities with data access and communication interfaces.
Minimal Architecture Extensions (MAEs): Functional augmentations of MAUs.
Minimal Architecture Patterns (MAPs): Higher-order orchestrations of MAUs/MAEs that generate full services.
Federated Architecture: The system is distributed across edge devices (laptops, servers) rather than monolithic clouds, ensuring IP protection, resilience, and low latency.

C. Research Approach

The authors utilize a bootstrapped Centre of Excellence (CoE) approach, combining:

Applied Science: Merging Economics, Technology, and Machine Learning.
Iterative Validation: Starting with the most complex domain (Algorithmic Trading) as a stress test, then abstracting the architecture to general corporate operations (Data MAPs).
Contrarian Strategy: Prioritizing "Exploration" (architectural perfection) over "Exploitation" (short-term revenue) to avoid the "spaghetti code" of legacy BigTech.

3. Key Contributions

The M1/M2 Distinction: A formal taxonomy separating the creation of models (M1) from the orchestration of intelligence (M2). The paper argues that the future of corporate value lies in M2, not M1.
Strategies-based vs. LLM-based Agentic AI: A clear delineation showing that while LLMs are useful for B2C information retrieval, they are structurally flawed for B2B transformation. True Agentic AI requires Strategies-based agents that combine heuristics with models.
The "Right-to-Win" Framework: A shift from "Right-to-Play" (using off-the-shelf tools) to "Right-to-Win" (proprietary, federated architectures). This repositions technology from a cost center to a source of competitive advantage.
Corporate AGI (CAGI): The introduction of a concept distinct from Human AGI. CAGI is a synthetic, federated intelligence designed for organizational efficiency, interconnectivity, and strategic alignment, achievable today without biological analogies.
The "95% Failure" Diagnosis: A structural explanation for AI project failures, attributing them to the mismatch between scientific methodologies (Science Applied) and business operational needs (Applied Science).

4. Results and Implementation

The authors detail a decade-long experimental journey resulting in two state-of-the-art platforms:

Fractal: An AI-first ERP platform that enables departmental transformation. It allows for the onboarding of AI products while simultaneously restructuring operational workflows.
AlphaDynamics: A platform for algorithmic trading and portfolio management that democratizes high-level financial strategies for non-experts (e.g., CFOs).

Key Outcomes:

Validation: The architecture has been validated in high-stakes environments (Algorithmic Trading, Cybersecurity, Banking) and audited by the Bank of Spain.
Scalability: The system successfully scaled from a single trading desk to a federated architecture capable of managing entire corporate groups and national economies ("Extreme-Efficient Nations").
Efficiency: The authors achieved a "Time-to-Production" (TTP) metric that outperforms traditional consulting and software development, proving that a small team (approx. 4.5 core individuals) can build complex, production-grade systems through architectural leverage.
Recognition: The work has received industry awards (Banking Tech Awards, CogX) and institutional interest from entities like the Bank of England, J.P. Morgan, and the Spanish Ministry of Economy.

5. Significance and Future Implications

Paradigm Shift: The paper challenges the prevailing narrative that "bigger models" equal "better AI." It posits that architectural discipline is the true bottleneck for AI adoption.
Economic Impact: By operationalizing Microeconomic theory through M2, the authors suggest a path to "Extreme-Efficient Nations" where GDP is enhanced through structural productivity gains rather than just model sophistication.
Defense and Security: The framework offers a new approach to "Novel Defense," where nations can defend against market manipulation and cyber-attacks using algorithmic counter-strategies rather than traditional kinetic or diplomatic means.
Academic Contribution: The paper argues for a new discipline of Applied Science that bridges the gap between academic theory and industrial execution, suggesting that current academic peer review is ill-suited for such interdisciplinary, high-risk innovation.
Future Roadmap: The authors outline a 20-year agenda to move from B2B transformation to B2C applications, the creation of "Orthogonal Art" (visualizing complex systems), and the establishment of new investment vehicles to fund this architectural evolution.

In conclusion, the paper asserts that Agentic AI is not an LLM phenomenon but a machine problem. Success depends on mastering M2 (Strategies-based Agentic AI)—the ability to build federated, self-orchestrating architectures that integrate human heuristics with machine intelligence, thereby transforming organizations from the inside out.